Skipping the Zeros in Diffusion Models for Sparse Data Generation

TL;DR

SED is a novel diffusion model that efficiently handles sparse data by modeling only non-zero values, saving computation and improving quality across various benchmarks.

Contribution

Introducing Sparsity-Exploiting Diffusion (SED), a method that preserves sparsity and reduces computation in diffusion models for sparse data.

Findings

SED matches or surpasses traditional DMs on benchmarks.

SED reduces unnecessary computation on zero entries.

Vision experiments highlight benefits of sparse modeling.

Abstract

Diffusion models (DMs) excel on dense continuous data, but are not designed for sparse continuous data. They do not model exact zeros that represent the deliberate absence of a signal. As a result, they erase sparsity patterns and perform unnecessary computation on mostly zero entries. With Sparsity-Exploiting Diffusion (SED), we model only non-zero values, preserving sparsity. SED delivers computational savings while maintaining or improving generation quality by skipping zeros during training and inference. Across physics and biology benchmarks, SED matches or surpasses conventional DMs and domain-specific baselines, while vision experiments provide intuitive insights into the limitations of dense DMs and the benefits of SED.